[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fV3ewT3kWWDSFMW-zpKUFSX5Z_VcfZm0tw2jfIWM9w4g":3,"$fdoG5HuuyISXeZniATKtBK8B7wTFlvSx4PY_50DKMn70":38},{"id":4,"name":5,"code":6,"logo":7,"cover":8,"url":9,"categoryId":10,"categoryName":11,"vendorId":12,"vendorName":13,"intro":14,"features":15,"pros":16,"cons":17,"pricingType":18,"priceNote":19,"freeQuota":20,"platform":21,"supportApi":22,"apiUrl":8,"score":23,"isRecommend":22,"isHot":22,"click":22,"sort":22,"createTime":24,"tags":25,"content":37},207,"Xiaomi-OCR-0","xiaomi-ocr-0","\u002Flogos\u002Fmimo.png",null,"https:\u002F\u002Fgithub.com\u002Fxiaomi-research\u002FXiaomi-OCR-0",9,"AI大模型",95,"小米","小米开源的 0.8B 轻量级多模态文档理解模型，文档解析、关键信息抽取与 OCR 视觉问答三合一","三能力合一：文档解析、关键信息抽取、OCR 视觉问答共享一套指令接口;0.8B 参数量，基于 Qwen3.5-0.8B，普通机器即可本地推理;Real5-OmniDocBench 95.24 分居首，覆盖扫描、弯曲、翻拍、光照、倾斜五种干扰;1.7 亿条训练样本由专家投票与失败驱动合成自动产出;完整开源：权重、推理代码与浏览器 Demo 一并放出;提供 MCP 服务与 Agent Skill 工具包，可直接接入智能体","0.8B 就做到 1B 以下文档解析第一，本地部署门槛低到可以忽略;三合一省掉多模型拼接的工程复杂度，接口统一;连 MCP 服务和 Agent Skill 都给了，接进现有系统基本是复制粘贴","只有文档理解能力，不做通用对话，不能替代通用多模态模型;综合榜单以 96.83 略低于阿里 TeleOCR 的 96.87;GitHub 与 Hugging Face 在国内访问需要额外的网络条件",1,"完全开源免费，权重与推理代码均已公开","无限制（本地或自部署）","本地部署\u002FWeb Demo",0,8.4,"2026-11-06 16:16:11",[26,30,33],{"id":27,"name":28,"color":29,"type":18},4,"开源","#909399",{"id":18,"name":31,"color":32,"type":18},"完全免费","#67C23A",{"id":34,"name":35,"color":36,"type":18},8,"国产自研","#F56C6C","Xiaomi-OCR-0 是小米开源的 0.8B 轻量级多模态文档理解模型，基于 Qwen3.5-0.8B 训练，把文档解析、关键信息抽取（KIE）与 OCR 视觉问答（OCR-VQA）三件事统一到一个模型里，通过共享指令接口切换任务，不需要再拼接多个专用模型。它的思路有点反直觉：不靠堆参数，靠数据。训练用了 1.7 亿条样本，由专家共识投票、渲染引导校验与失败驱动合成自动化产出，用数据质量换模型规模。效果上，官方给出的成绩是 OmniDocBench v1.6 拿到 96.83 分（第 2，仅次于阿里 TeleOCR 的 96.87），而在更贴近真实场景的 Real5-OmniDocBench 上以 95.24 分居首——这项基准专门覆盖扫描、纸张弯曲、屏幕翻拍、光照不均、倾斜拍摄五种干扰场景，小米在这五项全部领先。技术报告里还有一个有意思的发现：课程学习带来正向迁移，理解类监督在解析能力成熟之后反哺解析，最高提升 0.64 分，实现轻量模型下的双重能力跃升。开源程度相当彻底：不仅公开 0.8B 权重与推理代码，还提供了开箱即用的浏览器交互式 Demo、MCP 服务与 Agent Skill 工具包，开发者可以低成本把文档解析与视觉问答接进自己的业务流程或智能体系统。对需要在本地或私有环境处理票据、合同、表单的场景来说，0.8B 这个体量意味着普通机器就能跑，不必为文档理解单独买算力。",[39,44,46,51,53,58,60,65,69,73,77],{"id":40,"name":41,"color":42,"type":18,"toolCount":43},2,"免费增值","#409EFF",114,{"id":34,"name":35,"color":36,"type":18,"toolCount":45},69,{"id":47,"name":48,"color":49,"type":18,"toolCount":50},5,"支持API","#00B4A0",37,{"id":18,"name":31,"color":32,"type":18,"toolCount":52},32,{"id":54,"name":55,"color":56,"type":18,"toolCount":57},7,"移动端可用","#FF8A65",31,{"id":27,"name":28,"color":29,"type":18,"toolCount":59},15,{"id":61,"name":62,"color":63,"type":18,"toolCount":64},3,"付费使用","#E6A23C",12,{"id":66,"name":67,"color":68,"type":18,"toolCount":34},6,"可私有化","#7B61FF",{"id":70,"name":71,"color":72,"type":18,"toolCount":40},10,"长上下文","#3B55D9",{"id":74,"name":75,"color":76,"type":18,"toolCount":40},11,"多模态理解","#9C27B0",{"id":78,"name":79,"color":80,"type":40,"toolCount":52},16,"智能体","#10b981"]